Common Pitfalls In Reading Random Walks

Why Random-Walk Paths Can Be Misread

A sequence of values that rises over time can look like evidence of a predictable upward trend. But a random walk can rise for a while simply because its random increments happen to be mostly positive. Likewise, increments with a constant variance do not imply that the levels have a constant variance. Reading a random-walk plot correctly requires keeping track of what is random at each time: the level, the increment, or both.

This tutorial focuses on those distinctions and on interpreting first differences. The examples use hypothetical time series; they are not data from an insurer or another organization. The concepts are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Definition: A zero-drift random walk is a series $X_t$ defined by $X_t=X_{t-1}+\varepsilon_t$, where $X_0$ is fixed and the increments $\varepsilon_t$ are independent and identically distributed with mean 0 and finite variance $\sigma^2$. A random walk with drift $c$ is defined by $X_t=X_{t-1}+c+\varepsilon_t$, with increments having the same assumptions. The level $X_t$ is the value at time $t$; the increment is $X_t-X_{t-1}$.

The independence assumption is a convenient standard model for the examples below. The expectation and variance calculations also hold if the increments are uncorrelated, have the stated common mean and variance, and $X_0$ is fixed. An observed sequence, however, is just one realized path. Its shape need not match the expected path.

Trap 1: Treating a Rising Path as Deterministic

For a zero-drift random walk, the expected level is unchanged over time, but individual paths can move either upward or downward. An upward run is not proof that the next increment must be positive, nor does it establish a deterministic trend. A random walk with drift has a nonzero expected change per period, but its individual increments still include random variation.

Worked Example: A Rising Realized Path

A hypothetical index starts at $X_0=20$. Each period its increment is independently $+1$ or $-1$, each with probability $0.5$. The observed four increments are $+1,+1,+1,-1$. Find the levels, the expected level and variance at time 4, and the probability of this exact four-increment sequence. Does the observed path establish a deterministic rise?

Method and assumptions. This is a zero-drift random walk with $X_0=20$ and increments having probabilities $P(\varepsilon_t=1)=P(\varepsilon_t=-1)=0.5$. Each increment has mean $0$ and variance $1$.

Observed levels. Add each realized increment to the previous level:

X_1=20+1=21,\quad X_2=21+1=22,\quad X_3=22+1=23,\quad X_4=23-1=22.

The path rises for three periods and then falls. Its final value is $22$, which is 2 units above its start.

Expected level and variance. For a zero-drift random walk with fixed starting value, $E[X_4]=X_0=20$ and $\operatorname{Var}(X_4)=4\sigma^2=4$. The realized value $22$ differs from the expectation; that is not a contradiction. The expectation describes an average over possible paths, not a required endpoint for this path.

Probability of the exact sequence. Each of the four specified increments has probability $0.5$, and the increments are independent. Therefore, the probability of $+1,+1,+1,-1$ in that order is $0.5^4=1/16=0.0625$.

Conclusion and trap avoided. This path is one possible random outcome, not evidence that the increases were deterministic. The common mistake is to infer a fixed trend from a short upward run. The exact sequence has positive probability under the stated zero-drift model.

Trap 2: Assuming the Level Variance Is Constant

The increments in a random walk can have the same variance at every time while the levels become more variable as time passes. To see why, write a zero-drift level as the sum of its starting value and all increments up to that time. Each new increment adds uncertainty to the accumulated level.

Result: For a zero-drift random walk with fixed $X_0$ and uncorrelated increments of mean 0 and variance $\sigma^2$, $$ E[X_t]=X_0,\qquad \operatorname{Var}(X_t)=t\sigma^2. $$ For a random walk with drift $c$, the corresponding results are $E[X_t]=X_0+tc$ and $\operatorname{Var}(X_t)=t\sigma^2$.

The variance result follows directly from the definition. Repeated substitution gives $X_t=X_0+\sum_{j=1}^{t}\varepsilon_j$ in the zero-drift case. Because $X_0$ is fixed and the increments are uncorrelated, the variance of the sum is the sum of the increment variances. In the drift case, $tc$ is a constant and does not change variance.

\operatorname{Var}(X_t) =\operatorname{Var}\left(X_0+\sum_{j=1}^{t}\varepsilon_j\right) =\sum_{j=1}^{t}\operatorname{Var}(\varepsilon_j) =t\sigma^2.

Worked Example: Constant Increment Variance, Growing Level Variance

A hypothetical zero-drift index has fixed starting level $X_0=50$. Its independent increments are normally distributed as $N(0,4)$, where 4 is the variance. Find the mean and variance of the level at times 1, 5, and 12, and the standard deviation at time 12.

Method and assumptions. Apply the random-walk mean and variance formulas with $\sigma^2=4$. The normal distribution also means each $X_t$ is normally distributed, since it is a fixed value plus a sum of independent normal increments.

At time 1, $E[X_1]=50$ and $\operatorname{Var}(X_1)=1(4)=4$.

At time 5, $E[X_5]=50$ and $\operatorname{Var}(X_5)=5(4)=20$.

At time 12, $E[X_{12}]=50$ and $\operatorname{Var}(X_{12})=12(4)=48$. Its standard deviation is $\sqrt{48}$, approximately $6.9282$.

Check. The increment variance stays at 4 for every period, while the level variance increases from 4 to 20 to 48. The mistake to avoid is assigning variance 4 to every level because each new increment has variance 4. The level contains the accumulated increments, so its variance depends on the number of periods.

This increasing variance is one reason a random walk is not weakly stationary: its variance depends on time. Do not confuse that statement with saying that every realized path must spread smoothly outward. The variance describes the distribution across possible paths, not a guarantee about the appearance of any one path.

Trap 3: Confusing Levels With First Differences

The first difference of a series is the change from one time to the next. For a zero-drift random walk, this difference is exactly the increment. For a random walk with drift, it is the drift plus the random increment. Thus, the level and its first difference are different quantities with different behavior: levels accumulate changes, while differences record period-by-period changes.

\Delta X_t=X_t-X_{t-1} \quad\Longrightarrow\quad \begin{cases} \Delta X_t=\varepsilon_t & \text{for a zero-drift random walk},\\ \Delta X_t=c+\varepsilon_t & \text{for a random walk with drift }c. \end{cases}

Under the stated assumptions, the zero-drift first differences have mean 0 and variance $\sigma^2$. With drift, they have mean $c$ and variance $\sigma^2$. These statements concern the model distribution. A short sample of differences can have a nonzero sample average even when the model mean is zero.

Worked Example: Separating Levels, Drift, and Shocks

A hypothetical reserve index follows $X_t=X_{t-1}+1+\varepsilon_t$, with $X_0=40$. Assume independent increments $\varepsilon_t\sim N(0,4)$, so the drift is 1 unit per period and the shock variance is 4. The observed levels are $40,43,42,47,46$, at times 0 through 4. Find the observed first differences and implied shocks, their sample mean, and the model mean and variance of $\Delta X_t$.

Method and assumptions. Compute consecutive level changes first. Then subtract the drift from each change to recover the realized shocks. Do not treat the level values themselves as the increments.

First differences. In order, the changes are $43-40=3$, $42-43=-1$, $47-42=5$, and $46-47=-1$. Therefore, the observed differences are $3,-1,5,-1$.

Implied shocks. Since each difference is $1+\varepsilon_t$, subtract 1 from each observed difference. The shocks are $3-1=2$, $-1-1=-2$, $5-1=4$, and $-1-1=-2$.

Sample mean of differences. The sum is $3+(-1)+5+(-1)=6$, so their sample mean is $6/4=1.5$. This sample mean is not required to equal the model drift exactly.

Model properties and check. The model gives $E[\Delta X_t]=1+0=1$ and $\operatorname{Var}(\Delta X_t)=4$. The four observed differences average 1.5, while the four implied shocks average $(2-2+4-2)/4=0.5$; adding the drift of 1 gives the same sample difference mean, $1+0.5=1.5$.

Trap avoided. Reading the levels $40,43,42,47,46$ as though they were changes would give the wrong quantities. The first differences are the changes; with drift, subtracting the drift from those differences gives the shocks.

Conditional Forecast Variability

A related reading error is to use the variance of the level from the start of the series when the question is about uncertainty in a forecast from the current level. Suppose a zero-drift random walk is observed at time $t$, with $X_t$ known. Over the next $h$ periods, the forecast change is the sum of the next $h$ increments. Its expected value is 0 and its variance is $h\sigma^2$. Thus, the forecast mean of $X_{t+h}$, conditional on the observed current level, is $X_t$, and its conditional variance is $h\sigma^2$. With drift $c$, the conditional mean is $X_t+hc$, while the conditional variance remains $h\sigma^2$.

Worked Example: Forecasting From a Known Current Level

A hypothetical zero-drift random walk has independent increments $N(0,9)$. Its observed current level at time 8 is $X_8=100$. Find the conditional forecast mean, variance, and standard deviation of $X_{11}$, given the observed level at time 8.

Method and assumptions. The forecast horizon is $h=11-8=3$ periods. Use the three future increments, not all eight past increments, because $X_8$ is already known.

The conditional mean is $E[X_{11}\mid X_8=100]=100$, since the expected sum of the next three zero-mean increments is 0.

The conditional variance is $3(9)=27$, because the three future increments are independent and each has variance 9. The conditional standard deviation is $\sqrt{27}$, approximately $5.1962$.

Check and trap avoided. The forecast uses a three-period horizon. Using $11(9)=99$ would calculate the unconditional variance from a fixed starting point at time 0, not the conditional forecast variance given the known level at time 8.

Common Mistakes to Check

Calling a rising sample path deterministic. A realized rise can occur under a zero-drift random walk. Inspect the model assumptions before inferring a trend.
Keeping the level variance fixed. If increments have variance $\sigma^2$, then the zero-drift level variance from a fixed starting point is $t\sigma^2$, not $\sigma^2$.
Calling the levels the increments. Calculate $X_t-X_{t-1}$ to obtain first differences.
Ignoring drift in the differences. With drift $c$, the difference is $c+\varepsilon_t$, not just $\varepsilon_t$.
Mixing unconditional and conditional forecast variance. From a known current level, an $h$-period zero-drift forecast has variance $h\sigma^2$, not the variance calculated from the original starting time.

Key takeaway: A random-walk level accumulates increments. Its path can rise or fall by chance, its variance grows with elapsed time, and its first difference records the latest change (including any drift). Identify which of these quantities a question concerns before calculating.

Check Your Understanding

For each problem, identify whether the question concerns levels, increments, or first differences before calculating.

A zero-drift random walk starts at $X_0=5$. Its independent increments are $+2$ or $-2$, each with probability 0.5. The observed increments are $+2,+2,-2,+2$. Find the final level, $E[X_4]$, $\operatorname{Var}(X_4)$, and the probability of that exact sequence. (Answer: $X_4=9$, $E[X_4]=5$, $\operatorname{Var}(X_4)=16$, probability $=0.0625$.)
A random walk with drift $c=0.5$ starts at $X_0=12$, and its independent shocks have mean 0 and variance 9. Find $E[X_6]$, $\operatorname{Var}(X_6)$, and the standard deviation of $X_6$. (Answer: $E[X_6]=15$, $\operatorname{Var}(X_6)=54$, standard deviation $=\sqrt{54}\approx7.3485$.)
A series has levels $30,31,29,32$ at times 0 through 3. Find its three first differences and their sample mean. If these observations are treated as increments from a zero-drift random walk, does their sample mean establish nonzero drift? (Answer: First differences $1,-2,3$; sample mean $2/3$; no, a sample mean need not equal the model mean.)
A zero-drift random walk has independent increments with mean 0 and variance 2.5. Given $X_7=8$, find the conditional mean, variance, and standard deviation of $X_{10}$. (Answer: conditional mean $8$, variance $7.5$, standard deviation $\sqrt{7.5}\approx2.7386$.)
A zero-drift random walk starts at a fixed value and has uncorrelated increments with variance 9. A candidate says the level variance at time 5 is 9 because the increments have variance 9. Find the correct level variance and the variance of a one-period first difference. (Answer: Level variance $45$; first-difference variance $9$.)

What You'll Learn